精准的分布式光伏短期发电功率预测有助于电力系统运行与功率就地平衡.该文提出一种基于BIRCH(balanced iterative reducing and clustering using hierarchies)相似日聚类的L-Transformer(LSTM-Transformer)模型进行短期光伏功率预测.首先使用 BIRCH 无监督聚类算法对历史数据聚类得到 3 种典型天气,根据聚类结果划分测试集对模型进行训练.为提高不同天气类型下的预测精度,采用双层架构的L-Transformer模型,首层通过长短期记忆网络(long short term memory,LSTM)的门控单元机制捕捉时间序列中的长期依赖关系;次层结合Transformer模型的自注意力机制聚焦于当前任务更关键的特征量,通过多注意力头与光伏数据特征量相结合生成向量,注意力头并行计算,从而高效、精确地预测短期光伏功率.实测数据验证结果表明L-Transformer模型对于不同天气类型功率预测泛化性优异、精确度高,气象数据波动大时鲁棒性强.
Abstract
Accurate short-term power prediction of distributed photovoltaic systems is helpful to the operation of power systems and local power balancing.This paper proposes an LSTM-Transformer(L-Transformer)model based on BIRCH(balanced iterative reducing and clustering using hierarchies)for short-term photovoltaic power forecasting.First,the BIRCH unsupervised clustering algorithm is used to cluster historical data into three typical weather patterns,and the test set is divided according to the clustering results to train the model.To improve the prediction accuracy under different weather types,a two-layer L-Transformer model is adopted.The first layer captures long-term dependencies in the time series by using the gating unit mechanism of long short term memory(LSTM).The second layer combines the self-attention mechanism of the transformer model,focusing on the more critical features of the current task.By integrat-ing multiple attention heads with photovoltaic data feature quantities to generate vectors,the parallel computation of attention heads will efficiently and accurately predict short-term photovoltaic power.The results,combined with actual measurement data,show that the L-Transformer model has excellent generalization and high accuracy for power predic-tion under different weather types,and the robustness is strong when meteorological data fluctuations are remarkable.
关键词
深度学习/自注意力机制/多头注意力/BIRCH聚类/短期光伏功率预测/特征融合
Key words
deep learning/self attention mechanism/multi-head attention/BIRCH clustering/short-term photovoltaic power prediction/feature fusion
维普
万方数据